MRATNet: Learning Discriminative Features for Partial Discharge Pattern Recognition Via Transformers

Partial discharge pattern recognition (PDPR) is the fundamental cornerstone for fault diagnosis. It has emerged as a pivotal focal point in the field of power systems. However, PDPR faces several challenges, such as low signal quality and complex discharge patterns. We propose a multiscale residual aggregation transformer network (MRATNet) to address these challenges effectively. MRATNet learns long-dependent semantic relationships and discriminative features in partial discharge signals. Moreover, it integrates convolutional and transformer architectures as the feature extraction backbone. Thus, multiscale residual convolution blocks are incorporated to aggregate diverse information, and the transformer is leveraged to capture long-dependent semantic relationships. Meanwhile, the cross-attention mechanism is introduced to capture the spatial and channel feature distributions. The composite embedded feature selection module is proposed to extract discriminative features. Comprehensive experiments demonstrate the effectiveness of MRATNet, yielding exceptional performance on DEPD dataset (91.47%) and PDMDB dataset (86.05%). Finally, extended experiments have been conducted using the Technical University of Berlin's German emotional language library, suggesting the potential for generalizing our method to other recognition tasks.